The official implementation for ICCV 2025 CompressTracker General Compression Framework for Efficient Transformer Object Tracking.

[Models]

[Jun. 25, 2025]

• CompressTracker is accepted by ICCV 2025!

[Oct. 16, 2024]

• Code is available now!

[Sep. 28, 2022]

• We released our CompressTracker.

Our CompressTracker can be applied to any transformer tracking models. Moreover, CompressTracker supports any arbitrary levels of compression.

Our CompressTracker only needs an end-to-end and simple training instead of multi-stage distillation in MixFormerV2. The training cost is much lower than MixFormerV2.

Our compressTracker achieves better trade-off between performance and speed.

Option1: Use the Anaconda (CUDA 11.7)

conda create -n compresstracker python=3.9
conda activate compresstracker
bash install.sh

Option2: Use the Anaconda (CUDA 11.7)

conda env create -f compresstracker_env.yaml

Run the following command to set paths for this project

python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir ./output

After running this command, you can also modify paths by editing these two files

lib/train/admin/local.py  # paths about training
lib/test/evaluation/local.py  # paths about testing

Put the tracking datasets in ./data. It should look like this:

${PROJECT_ROOT}
 -- data
     -- lasot
         |-- airplane
         |-- basketball
         |-- bear
         ...
     -- got10k
         |-- test
         |-- train
         |-- val
     -- coco
         |-- annotations
         |-- images
     -- trackingnet
         |-- TRAIN_0
         |-- TRAIN_1
         ...
         |-- TRAIN_11
         |-- TEST

Download pre-trained MAE ViT-Base weights and put it under $PROJECT_ROOT$/pretrained_models. Besides, please download the pretrained OSTrack weights and put it under $PROJECT_ROOT$/pretrained_models, too.

python tracking/train.py --script compresstracker --config compresstracker_vitb_256_4 --save_dir ./output --mode multiple --nproc_per_node 8

Replace --config with the desired model config under experiments/compresstracker.

It is worth noting that our CompressTracker can support any structure, any resolution, and any level of compression. We provide the code for CompressTracker-2/3/4/6 in our paper, and you can easily modify it to compress your own model.

Download the model weights from Google Drive(Maybe not the same as the reported performance in paper)

Put the downloaded weights on $PROJECT_ROOT$/output/checkpoints/train/compresstracker

Change the corresponding values of lib/test/evaluation/local.py to the actual benchmark saving paths

Some testing examples:

• LaSOT or other off-line evaluated benchmarks (modify --dataset correspondingly)

python tracking/test.py compresstracker compresstracker_vitb_256_4 --dataset lasot --threads 64 --num_gpus 8
python tracking/analysis_results.py # need to modify tracker configs and names

• TrackingNet

python tracking/test.py compresstracker compresstracker_vitb_256_4 --dataset trackingnet --threads 64 --num_gpus 8
python lib/test/utils/transform_trackingnet.py --tracker_name compresstracker --cfg_name compresstracker_vitb_256_4

Note: The speeds reported in our paper were tested on a single RTX2080Ti GPU.

# Profiling vitb_256_mae_ce_32x4_ep300
python tracking/profile_model.py --script compresstracker --config compresstracker_vitb_256_4

• Thanks for the OSTrack, PyTracking library, MixFormerV2, and OneTracker, which helps us to quickly implement our ideas.
• We use the implementation of the ViT from the Timm repo.

If our work is useful for your research, please consider citing:

@article{hong2024general,
  title={General Compression Framework for Efficient Transformer Object Tracking},
  author={Hong, Lingyi and Li, Jinglun and Zhou, Xinyu and Yan, Shilin and Guo, Pinxue and Jiang, Kaixun and Chen, Zhaoyu and Gao, Shuyong and Zhang, Wei and Lu, Hong and others},
  journal={arXiv preprint arXiv:2409.17564},
  year={2024}
}